Source Modeling of Severely Pathological Voices - EE @ UCLA

Source Modeling of Severely Pathological Voices

Bruce R, Gerratt*, Jody Kreiman*, Norma Antonanzas-Barroso*, Brian Gabeiman*, and Abeer Alwan* * “Division of Head and Neck Surge~

and *“Dept. of Electrical Engineering,

UCLA, Los Angeles, CA 90095

USA

Abstract: Our previous attempts at synthesizing severely pathological voices were hampered by perceptuallysignificant errors in vowel quality. To determine whether these were caused by methodological limitations or by source-tract interactions, normai and pathological voices were recorded using both a condenser microphone and a flow mask system. Synthetic versions of each signal were produced and compared perceptually to the original signais. Limitations of all-zero inverse filtering techniques as appiied to pathological voices wiii be discussed.

INTRODUCTION

Attempts at modeiing pathological phonation using anaiysis by synthesis depend on accurate modeling of the pathological voice source and vocai tract resonances, The source finction is typically estimated using inverse filtering, either of the glottal flow signal (e.g., 1) or of the acoustic signai as transduced with a condenser microphone (2). Because of the restricted frequency response of the flow mask system, our preliminary experiments synthesizing pathological phonation used estimates of vocal tract characteristics derived from a microphone signal, combined with estimates of vocai tract resonances derived from a flow mask signal (recorded from a different segment of the same utterance). These studies suggested that the frequency response characteristics of the flow mask sys~em may limit its applicability to severely pathological phonation, particularly when precise source modeiing is required. In addition, perceptually-salient changes in vowel quality occurred when we resynthesized voices using formants estimated from the microphone signal and sources estimated from the flow signal. These difficulties may be due to the fact that microphone and flow signals were measured at different points in an utterance, or to the fact that significant source-fiiter interactions occur in pathological voices (or both). To assess the extent to which flow and acoustic signais differ in their abiiity to capture source and vocal tract characteristics ofpathoiogical phonation, the foiiowing experiment was undertaken.

RECORDING

~THOD

Eight speakers participated in this experiment. Four (two males, two females) were randomly-selected patients complaining of voice disorders; four (two maies, two females) were volunteers free from vocai pathoiogy. For haif the subjects, a flow mask (Giottai Enterprises) was piaced over the subject’s face and a 1” condenser microphone (Bruel & Kjaer) was held a constant distance from the subject’s lips outside the flow mask. The subject sustained the vowel /d for at ieast 2 seconds. Approximately haIf way through the utterance, the flow mask was quickiy removed, aiiowing recording of the acoustic signal without any mask effects. The order of recording was reversed for the remaining subjects, with acoustic signais recorded first and flow mask appi ied half-way through the utterance. Thus, both flow and acoustic signais were recorded from a single utterance for each subject. Ali signais were iow-pass fiitered at 8 kHz and sampied at 20 kHz.

1271

ANALYSIS

AND SYNTHESIS

TECHNIQUES

Signals were down-sampled to 10 kHz prior to analysis. Two sets of analyses were undertaken. The first (flow mask-microphone condition) used the flow signal to estimate the source function, and the microphone signal to estimate the vocal tract response. To derive the glottal pulse shape from the flow signals, forrnants and bandwidths for inverse filtering were estimated as follows. The beginning of the closed phase was located through LPC error analysis, and a covariance LPC analysis was performed (30 to 40 samples, order 12 or 14). If this procedure failed, or if there was no closed phase, an autocomelation LPC analysis was performed instead (256 points, order 12 or 14). The flow signal was then inverse filtered using the all-zero filter method described in (3). In the second condition (microphone only condition), inverse filtering was performed directly on the acoustic signal recorded with the condenser microphone. Formants and bandwidths for inverse filtering were estimated as described above, but from the acoustic recording. In both conditions, formants and bandwidths were manipulated interactively during inverse filtering to produce the “best” result possible. Our primary criterion for success was a smoothly decreasing source spectrum slope. Formant frequencies and bandwidths for synthesis were estimated in both conditions from the microphone signals using autocorrelation LPC (256 points, order 12 or 14), supplemented with spectrographic analysis. Note that for the flow mask-microphone condition, the source and formant analyses were performed on different segments of a single utterance. For the microphone-only condition, both analyses were performed on the same segment of speech. Three different versions of each voice were synthesized for each condition: a single pulse extracted from the original voice signal and concatenated to form a 1 second stimulus to avoid the perceptual effects of vocal perturbation and other sources of noise, a “reconstructed” version comprising the source function produced from inverse filtering (of the flow signal in the flow mask-microphone condition, or of the microphone signal in the microphone only condition) and the vocal tract resonances estimated from the microphone signal; and an “LFfitted” synthesized version created by combining the estimated vocal tract resonances with a source that had been least-square fitted with a simplified LF source model (4).

PERCEPTUAL

EVALUATION

Ten expert listeners assessed the similarity of each reconstructed and synthetic stimulus to the original (natural) voice on a visual analog scale. Perceptual analyses will compare the success of modeling normal and pathologic phonation and male and female voices, as well as the relative success of modeling efforts using the

flow mask-microphone and condenser microphone only techniques.

ACKNOWEDG~NT

This research was supported by N~CD grant DCO 1797.

RE~RENCES

1. Rothenberg,M., J Acoust,Sot. Am. 53, 1632-1645(1973). 2. Ananthapadmanabh~T.V., STL-QPSR 2/3, I-24(1984). 3. Javkin,H., Antonanzas-Barroso,N., and Maddieson,1.,J Speech Hear. Res. 30, 122-129(1987) 4. Qi, Y., and Bi, N., J Acoust. Sot. Am. 96, 1182-1185 (1994).

1272